Advanced building detection in VHR satellite imagery:
a comprehensive study using different mask R-CNN approaches

Advanced building detection in VHR satellite imagery: a comprehensive study using different mask R-CNN approaches

From 16 to 19 September 2024, the symposium was held in Edinburgh Artificial Intelligence and Image and Signal Processing for Remote Sensing XXX.

Each year, this important community gathers to share and discuss current research, hear about the latest findings, network with colleagues, and to explore topics on the various research areas in Security, Defence and Remote Sensing.

Topics include research into satellite monitoring of the atmosphere and imaging of terrestrial ecosystems, as well as research into sensors and photonic technologies for security and defence.

Also on the agenda this year is the Paper Advanced building detection in VHR satellite imagery: a comprehensive study using different mask R-CNN approacheswhich managed to pass the very tough selection thanks to the painstaking and innovative work of its authors:Luca Galli, Martina Infante, Edoardo Unali, Alberto Gallottini di Exprivia SpA.

The accurate identification of building footprints from high-resolution satellite images is crucial for urban planning and disaster response. 

This paper analyses building detection methodologies using the Mask R-CNN framework and its variants, with the aim of addressing challenges such as accurate classification of boundary pixels and reduction of false positives.

Two WorldView-3 datasets are used for the analysis, including the SpaceNet building detection dataset and a dataset on Prato, Italy. Augmented techniques, such as NDVI and Sobel edge detection features, and evaluation metrics such as F1 score and average accuracy are used to evaluate the performance of the model.

The results reveal the superiority of the Point Rend Mask R-CNN in detecting smaller buildings in densely populated urban environments, while the RGB Cascade B Mask R-CNN performs well on the Lawn dataset. In particular, the Point Rend and Cascade models demonstrate substantial improvements over other building detection methods. This investigation provides insights into the effectiveness of the Mask R-CNN framework and its variants to advance building footprint delineation in various applications.